SBC seminars 2006

This talk will conclude with a brief overview of some of Prof. Casavant's other research areas, including some applications of Bioinformatics to Medical Genetics in disease gene mutation finding.

• The Human Metabolome Project
• The Human Metabolome Database
• Protein-Based Drug Targeting and Medical Diagnostics: the development of new methods for the targeted delivery of drugs or diagnostic agents using peptide or protein-based vehicles;
• NMR Technology Development: the development novel applications in NMR spectroscopy to facilitate structure-aided drug design;
• Pharmaceutical Bioinformatics: the development of innovative bioinformatics and modeling software for improved protein/peptide analysis.

Background

Results

Conclusions

Contents:
- Protein prediction networks using the STRING database
- Domain co-occurrence networks using the SMART database
(- Networks of transcription during Drosophila embryogenesis)
	    

Outline:

1. Introduction to kernel methods.
1.1. Basic concepts: Learning and similarities.
1.2. Intuitive concept of kernels.
1.3. Kernel trick, Mercer's kernels, and the Representer Theorem. 
1.4. Linear methods again through positive definite kernels. 
1.5. Properties of kernels.

2. Support Vector Classifier (SVM).
2.1. The optimal separating hyperplane.
2.2. Maximum margin classifiers
2.3. Optimization.
2.4. Non-separable problems.
2.5. Hands on training

3. Support Vector Regression (SVR).
3.1. Formulation, the tube, and the kernel trick.
3.2. Tuning the SVR with profiled insensitivity.
3.3. Non-linear regression.
3.4. Hands on training.

4. Designing the kernel.
4.1. Incorporating prior knowledge in the kernel.
4.2. Kernel engineering.
4.3. The profile-dependent SVM.
4.4. Kernelizing a linear method.

5. Examples on Bioinformatics.
5.1. Kernels for protein sequences.
5.2. Kernels in AO and siRNA efficacy prediction.

6. Further information.
6.1. Demos.
6.2. Source Code.
6.3. Useful links.

7. Conclusions and trends.

Domain architectures are not necessarily created once only during evolution. Cases of convergent evolution show how a favourable domain architecture has evolved multiple times independently. A basic concept for understanding evolution on gene level is orthology.

Two genes are orthologous if they have evolved from the same gene in the last common ancestor of the species and have thus been created by a speciation event. Paralogous genes result from a duplication event that produced two gene copies within the same species.The concept of orthology can be transferred from genes to protein domains and utilised to explain recombination of protein domains and the evolution of domain architectures.

The focus of this work is to augment the understanding of domain architecture evolution and its functional implications. We have examined, evaluated and improved existing methods as well as developed new approaches. The concept of orthology plays a major role in this work. Orthology is often inferred from phylogenetic trees that are based on pairwise distance estimations of protein sequences. The Scoredist protein sequence dis- tance estimator has been developed as one part of this thesis. It combines robustness with low computational complexity and can be calibrated towards various evolutionary models.

Accurate phylogenetic trees are crucial for many applications, hence the appropriate tree reconstruction algorithm should be chosen with care. The strengths and weaknesses of many current tree reconstruction algorithms were assessed, and findings underscore the value of the Scoredist estimator. The Pfam protein families database comprises a large number of protein families and domains. As part of this thesis it has been enhanced by search and query tools, such as PfamAlyzer or the browser-based domain query, that can be applied on whole domain architectures instead of individual domains only.We have developed a Maximum Parsimony algorithm for the prediction of ancestral domain architectures. In contrast to previous approaches, it employs gene trees rather than species trees. The algorithm was a starting point for an extensive study of the do- main architectures present in Pfam for 50 completely sequenced species. Sampling widely across the kingdoms of life, the study sought to find and analyse cases where a domain architecture had been created multiple times. The algorithm proved robust to potential biases from horizontal gene transfer. Convergent evolution of domain architectures was found more frequently than by previous approaches. No strong biases driving convergent evolution were found. It therefore seems to be a random process in much the same way evolution through duplication and recombination, yet less frequent.

These domains (and motifs) are more conserved during evolution and constitute the basic elements of modular evolution within and in between organisms. Most proteins consist of two or more domains, giving rise to a variety of combinations of domain arrangements. Several events can be more easily traced by viewing proteins as strings of domains rather than as strings of amino-acids. For example, circular permutations represent non-linear arrangements of domains (e.g., the string of domains ABCD and CDAB are circularly permutated with respect to each other) such that the sequences they can not be detected using the standard sequence alignment procedures which depend on the recursive scheme but the strings of domains can be easily analysed bacause of the smaller search space. We have analysed the modular evolution of proteins and found that the major events creating novel arrangements are fusion and loss of domains at the ends and, to a lesser extend fission. Recombination seems to play only a minor role.

In the group has previously been developed MCMC algorithms for integrated modeling of genome evolution where the rates are instead drawn independently and \ identically from an underlying distribution of choice. In the presented work we have made hill-climbing maximum likelihood (ML) variants of the above algor\ ithms and I will show that the ML approach is enabling accurate simultaneous inference of rates and times on large trees.

Results: In proteins with signal peptides, there is an overabundance of phase 1 introns around the region of the signal peptide cleavage site. This has been described before. But in proteins without signal peptides, a novel phenomenon is observed: There is a sharp peak of phase 0 intron positions immediately following the start codon, i.e. between codons 1 and 2. This effect is seen in a wide range of eukaryotes: Vertebrates, arthropods, fungi, and flowering plants. Proteins carrying this start codon intron are found to comprise a special class of relatively short, lysine-rich and conserved proteins with an overrepresentation of ribosomal proteins. In addition, there is a peak of phase 0 introns at position 5 in Drosophila genes with signal peptides, predominantly representing cuticle proteins.

Conclusion: There is an overabundance of phase 0 introns immediately after the start codon in eukaryotic genes, which has been described before only for human ribosomal proteins. We give a detailed description of these start codon introns and the proteins that contain them.

FunCoup (http://www.sbc.su.se/~andale/funcoup.html) is a statistical framework of data integration for finding functional coupling (FC) between proteins. It is capable of transferring information from model organisms via orthologs found by InParanoid program. Data of different sources and various natures (contacts of whole proteins and individual domains, mRNA co-expression, protein co-occurrence in tissues and cellular compartments, similar phylogenetic profiles etc.) are collected and probabilistically evaluated in a Bayesian network (BN), trained on setsof known FC cases vs. sets of randomly picked protein pairs as background reference. To address known drawbacks of Bayesian estimators, FunCoup was optimized in several aspects and, compared to previous framework configurations of this sort, the net gain in performance is tens of percentage points in either sensitivity or specificity. The number of simultaneously used model organisms (5-8) and individual datasets (50-60) has been estimated as maximal for practical purposes. It means that no further significant gain is expected given the current state of high-throughput data and the standard Bayesian framework.

We are, however, developing a post-processor where evidence that supports individual links is cross-validated by agreement/interaction between data types, model organisms, and network context that two proteins share. Combinations that efficiently augment the FC discovery rate are accepted as terms in a regression model. The model is then used to correct the scores found by the FunCoup BN. The predicted functional links thus become even more confident.

FunCoup is a self-consistent framework that easily incorporates nearly any kind of data (continuous values of any distribution shape, binary data, character labels etc.) from any data source without human curation. It has thus been possible to generate networks for several organisms in respect of different types of functional coupling. A network for Ciona intestinalis, which had neither training sets nor sources of its own data, was created as well.

Since the last year SBC seminar talk, the networks for human, mouse, rat, worm, fly, Ciona, Arabidopsis, and yeast have been made available on the FunCoup website, which has also acquired a spectrum of visual (due to Medusa network applet - Hooper and Bork, 2005) and download functionalities. Each link in FunCoup contains information about underlying evidences and a confidence value.

In the last ten years, knowledge about the function of many proteins has been gathered through high-throughput studies of protein-protein interactions (PPI). As a result, proteins of virtually unknown functions found in complexes with a number of proteins of known, and related, function may be tentatively associated with the functionality of its interacting partners. In the beginning of this year two independent studies of genome-wide screens for the complexes of the S. cerevisiae proteome using tandem affinity purifications (TAP) were published. Although the two TAP methods are very similar, the resulting complexes are clearly different. Hence, inferring function from them may render varying results depending on the dataset used.

We have sought to investigate the differences between the various datasets for S. cerevisiae PPI and assess their reliability compared to literature curated datasets. Further, we aim to understand if distinct methods may detect different types of interactions and if such biases can be contributed to specific protein properties. It has been shown that proteins that interacts with many partners (hubs) often are highly conserved and essential. In addition, proteins that interact are often co-expressed. Other properties of hub proteins have been suggested, such as high content of disordered structure and high content of repeating domains. We find that these properties seem too be especially important for interactions that are detected with yeast two hybrid screens. Further, we show that abundant proteins are common in networks derived from TAP studies.

Knowledge and identification of protein-protein interactions and exact antibody epitopes is of great importance for many antibody applications, in particular antibodies aimed at clinical usage.
Here we describe a novel high throughput method for identification of antibody binding epitopes. Currently we produce and screen target libraries of approximately 10^4- 10^5 peptides per antibody target, with an epitope mapping resolution down to 7aa.
The technique involves fragmentation of target DNA and subsequent blunt-end ligation into a gram-positive expression vector. The resulting peptide library is electroporated into Staphylococcus carnosus bacteria and transported for epitope presentation as a membrane anchored fusion protein on the cell surface. Cells are incubated with corresponding antibody and analysed using Fluorescence Activated Cell Sorting (FACS). Both positive cells, presenting epitopes recognised by the antibodies, as well as negative, non-epitopes, are sorted out separately in a 96 well format. These epitope and nonepitope cells are grown and sequenced using Pyrosequencing. Sequence data is interpreted and epitope binding sites mapped back to the target protein sequence using BLASTP and BioPerl.
Our current results include mapping of monoclonal, polyclonal and monospecific antibodies targeting HER2. Our ongoing study includes a large scale monitoring of antibody immune response in rabbits, focusing on identification and prediction of antigenic motifs, using HPR antibodies (www.hpr.se and www.proteinatlas.org).